How To Write With An LLM

Simon Willison highlights Thomas Ptacek’s blunt rule for writing with LLMs: treat them as copyeditors, not as writing partners. The line is simple enough to stick in your head and strict enough to matter. If an LLM suggests a specific turn of phrase, that phrase is off limits. No exceptions, no little compromises, no “just this once.”

That sounds harsh until you hear the reason. Ptacek says he uses LLMs for fact-checking, spelling, grammar, and the occasional thesaurus, but not for drafting blog content. The point is to keep the machine from sneaking its own voice into the text. He describes that voice as having a weird smell, which is as good a complaint as any about model-generated prose.

Willison says the rule feels like a kind of intellectual personal protective equipment, and that’s the right way to read it. Not as anti-AI purism. More like a guardrail for people who want the speed and utility of these tools without letting them quietly flatten their writing into the model’s habits. There’s a real discipline in refusing the first polished phrase that shows up.

The piece also points to Ptacek’s own LLM copyediting tool, shown in a screenshot, plus a prompt he shared for building something similar. So the message isn’t “don’t use LLMs.” It’s use them with a hard boundary, and keep the final voice human, even if the machine helped clean up the commas on the way there.